Meta Removes Hidden Chatbot Instructions Exposed on WhatsApp Help Pages
A bug on WhatsApp’s support site leaked internal directives for Meta’s AI assistant, including rules on what not to say to users.
Key points
- Around 80 WhatsApp help pages exposed hidden code blocks containing instructions for Meta’s chatbot.
- The leaked text included directives such as "WHAT NOT TO SAY" and advice against recommending regulatory contact.
- Meta confirmed a bug caused the exposure and stated the issue has been fixed after media inquiry.
Meta removed hidden instructions from approximately 80 WhatsApp help pages after a privacy firm discovered they were exposed in the website’s code. These directives, intended for Meta’s internal chatbot, included specific guidance on how the AI should interact with users seeking support.
The exposed text appeared within code blocks labeled “HelpContentGalaxyLLMInclude” and contained written details not visible to standard users. According to findings from privacy company Redact, the instructions provided steering for the large language model to avoid unhelpful responses. One example on the account deletion page advised the bot not to recommend contacting data protection authorities or regulators.
Context and Background
Dan Saltman, CEO and founder of Redact.dev, noted the disparity between public-facing advice and the internal commands given to the AI. His team regularly scans tech firms’ sites and privacy policies for such discrepancies. Saltman stated that while many instructions were straightforward, pointing users to official support channels, they were clearly not intended for public exposure.
WIRED Security reported that the hidden text could still be viewed in the underlying code of archived versions of the help pages. The analysis found about 100 blocks of hidden text across the site. This exposure offered a glimpse into how Meta creates instructions for its customer service automation tools.
Who Is Affected
The incident primarily affects WhatsApp users who might have accessed these help pages during the window the bug was active. However, the text was embedded in code rather than displayed prominently, limiting casual visibility. It remains unclear if third-party chatbots could access this text or how long the instructions had been publicly visible before removal.
After WIRED contacted Meta about the issue, the company acted to remove the text from all affected pages. WhatsApp spokesperson Joshua Breckman confirmed that a bug caused the display of these internal instructions. He emphasized that the text was meant for the chatbot to ensure accurate responses for users messaging WhatsApp Support with AI.
What should you do?
Security professionals and privacy advocates should monitor archived versions of tech company help pages for similar exposures. Organizations using AI for customer support should audit their public-facing code to ensure internal prompts are not inadvertently accessible. Users encountering unusual guidance on support pages should report it to the vendor immediately.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Meta
- Review archived versions of vendor help pages to check for exposed internal code or prompts.
- Audit public-facing web properties to ensure AI system instructions are not rendered in visible code blocks.
- Monitor vendor communications for bug reports related to accidental data or instruction exposure.
- Report any discovered internal directives or hidden text to the responsible vendor’s security team.
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What kind of instructions were leaked on WhatsApp's website?
The leaked instructions were directives for Meta's chatbot, including sections titled "WHAT NOT TO SAY" and advice on avoiding regulatory recommendations.
How many pages were affected by this bug?
According to Redact, around 80 help and FAQ pages across WhatsApp’s website exposed the hidden text instructions.
Has Meta fixed the issue?
Yes, a WhatsApp spokesperson confirmed that a bug caused the issue and that the text has been removed from all pages.
